Your brain estimates how often things happen by how quickly examples show up — a decent proxy, until the news, your feed, and your own vivid memories start writing to the cache. This chapter is about what happens when retrieval speed impersonates statistics.
A week of plane-crash headlines, and suddenly the drive to the airport feels safer than the flight. Nothing about aviation changed. Your cache did.
Ask yourself "how dangerous is flying?" and watch what actually executes. You don't consult a table of fatalities per passenger-mile — you don't have one. What you have is a memory, and memory answers a different call: fetch me crashes. Three arrive instantly, in color, with cockpit audio. Request complete. Verdict: dangerous.
This is Chapter 3's substitution running on a specific target. The hard question — how frequent is it? — gets silently swapped for an easy one — how easily do examples come to mind? — and the answer to the second is returned as the answer to the first, no cast, no warning, no deprecation notice. Kahneman and Tversky named the swap the availability heuristic.
As heuristics go, it isn't stupid. Frequent things really do leave more traces, so retrieval latency genuinely correlates with frequency — judging counts by cache-hit speed is a defensible O(1) approximation. It fails the way every cache fails: when something other than frequency did the writing. Vividness writes. Recency writes. Personal involvement writes. And one process writes harder than everything else combined — a news industry whose entire selection function is "unusual enough to be interesting." Your sample of the world would need to be representative for this heuristic to be safe. It almost never is.
Book the drive. Flying feels dangerous right now — three vivid crashes came to mind before you finished the thought. That's not caution talking, that's evidence: if disasters are this easy to retrieve, they must be everywhere. Twelve hours on the interstate suddenly reads as the prudent option.
Ease of recall is measuring the news cycle, not the hazard. Three crashes came to mind because three crashes were broadcast — the denominator (about a hundred thousand uneventful flights a day) never airs. Per mile, the drive is the risky choice by orders of magnitude; swapping a flight for a long road trip raises your actual fatality risk while lowering the felt one.
The transferable habit: when a risk feels suddenly bigger, ask what changed — the world, or your feed? If the answer is "my feed," the feeling is telemetry about the sampler, not the hazard.
Consider the letter K. Is it more common as the first letter of an English word, or the third?
Most people say first, and quickly. Watch why. Asked for K-words, you generate kitchen, kangaroo, kite before the question mark lands. Asked for third-letter-K words — _ _ k … — you get… ask? acknowledge? The retrieval grinds. First position feels more common because first position is indexed: your mental lexicon supports lookup-by-prefix and nothing remotely like lookup-by-third-character. The reality: in typical English text, K appears roughly twice as often in third position as in first.
The wrong answer isn't a fact about English. It's a fact about your index structure, leaking. Judging letter frequencies by retrieval speed is like judging row counts by query latency when one column has an index and the other forces a full table scan — the latencies differ by orders of magnitude; the row counts needn't differ at all.
Ask each member of a couple, separately, what percentage of the housework they personally do. Add the two numbers. They reliably total more than 100.
That's the classic Ross and Sicoly result, and it is not (only) spin. You were present for every dish you washed, every 2 a.m. bottle, every bin you hauled out — your attendance rate at your own contributions is exactly 100%. Your partner's contributions you caught only when you happened to be in the room. Both people query the same kind of biased cache, both get inflated self-counts, and both are sincere. That's the nasty part: an availability bias produces the same output as self-serving dishonesty while feeling, from the inside, like plain observation.
The same arithmetic runs on any team. Everyone remembers their own incident heroics, their own late-night reviews, in full resolution; everyone else's are a lossy sample. Co-authors each privately believe they carried the paper — Kahneman notes the total claimed credit on joint projects routinely exceeds 100%. The effect even runs on negatives: people also over-claim their share of the arguments they started. It's not vanity, it's indexing — the cache is self-keyed.
List six times you acted assertively, then rate how assertive a person you are. Now imagine we'd asked for twelve instead.
More evidence should mean a stronger conclusion — twelve examples beat six. That is not what happens. In Norbert Schwarz's experiment, people who listed twelve rated themselves less assertive than people who listed six. The first six flowed; items seven through twelve had to be dredged, and the dredging itself was entered into evidence: if I'm struggling this hard to find examples, there must not be many. The struggle outvoted the pile.
So the heuristic is subtler than "count what comes to mind." What System 1 actually monitors is the ease of retrieval — and the feeling of ease can trump the content retrieved. Best of all is the failure mode that proves the mechanism: give people an excuse for the difficulty — background music that supposedly interferes with recall — and the effect flips. With the strain explained away, they fall back on the content, and the twelve-listers feel more assertive again. The feeling only counts as evidence while it's unexplained. Attribute it to anything else, and System 1 quietly drops it from the record.
Which kills more people in a typical year — tornadoes, or asthma? Calibrate yourself before reading on; the widget keeps score.
In the classic risk-perception studies (Lichtenstein, Slovic and colleagues, 1970s), people judged tornadoes deadlier than asthma — asthma killed on the order of twenty times more. Accidents were judged over 300 times deadlier than diabetes; the true ratio ran the other way, about four to one in diabetes' favor. The pattern was systematic: dramatic, story-shaped deaths — the ones with wreckage and footage — overweighted; quiet, chronic ones nearly invisible. And the estimates tracked one variable beautifully: media coverage. The news covers what's novel and vivid, because that is what news is. Your risk model was trained on their sampler, and the sampler's selection bias is the whole product. Substitute "engagement-ranked feed" for "evening news" and the 2026 version has a bigger amplifier and a personalization layer.
Slovic pushed it one step darker. Ask people about a technology — pesticides, nuclear power — and their risk and benefit estimates come back negatively correlated, which reality's are not (high-benefit things are often high-risk; that's why we bother with them). The judgment running underneath is simpler: do I like it? Tell people a technology's benefits are high, and their estimate of its risk drops — with no new risk data supplied. Likes and dislikes come first; the risk numbers are backfilled to match. Slovic called it the affect heuristic: availability's sibling, where the thing that "comes to mind" isn't an example but a feeling.
Love Canal, 1978: buried chemical waste surfaces in a New York neighborhood. The health damage was real but contested. The coverage was not.
Watch the mechanism, because it's a feedback loop and every stage is individually reasonable. An incident makes vivid news. The coverage makes the risk available, so the public worries. Worried publics are an audience, so coverage increases. Rising volume reads as rising importance, so politicians respond — and a political response is itself news. Around again. A few cycles in, the emotion is the story, anyone questioning the panic is suspected of cover-up, and the eventual policy is proportional to the volume of the loop, not the size of the hazard. Kuran and Sunstein named the pattern the availability cascade.
The Alar apple scare of 1989 is the type specimen. A TV segment calls an apple-ripening chemical "the most potent carcinogen in the food supply" — an extrapolation from enormous rodent doses — and the loop engages: panic, apple juice poured down drains, an industry out more than $100 million, all for a risk that later review put somewhere between tiny and unmeasurable. Meanwhile asthma, which reliably kills thousands every year, has never once trended.
It's tempting to pick a side — trust the experts, or trust the public's fear. Kahneman declines, and he's right to. The public sometimes smells what experts dismiss: Love Canal helped put toxic-waste cleanup law on the books, and few regret that. Experts anchor on expected fatalities; citizens also weigh dread, voluntariness, fairness — and those aren't obviously illegitimate weights. But whatever your risk portfolio should optimize, "whatever was loudest this quarter" is not it. Both errors are real, and they share a cause: availability is a terrible prioritizer. It funds the vivid and starves the merely lethal.